Security Review Skill

Comprehensive security checklist for full-stack applications.

When to Use This Skill

• Reviewing API endpoints for vulnerabilities
• Auditing frontend code for XSS/CSRF risks
• Validating authentication and authorization patterns
• Checking data handling and storage security
• Pre-merge security review
• Reviewing AI-generated code for common pitfalls
• Evaluating MCP server configurations for security risks

Security Checklists

AI-Assisted Development Risks

AI-ism Detection Checklist

When reviewing AI-generated code, specifically check for these common LLM output patterns:

1. Verify all imports: Run npm ls <package> for any unfamiliar package. Check npm for download count (>10k weekly) and last publish date
2. Check for invented APIs: LLMs sometimes generate method calls that look plausible but don't exist (e.g., prisma.task.findFirstOrCreate())
3. Scan for insecure defaults: Look for permissive CORS, disabled auth middleware, any types masking validation gaps
4. Watch for training data leaks: Generic variable names, TODO comments from other projects, or boilerplate that doesn't match project conventions
5. Validate error handling: AI often generates optimistic code paths without proper error handling or with catch blocks that swallow errors silently

MCP Server Security

API Endpoints (Express/Prisma)

Frontend (Vue/Pinia)

Authentication & Authorization

Data Handling

Output Format

When reporting security findings, use this structure:

## Security Review: [File/Feature Name]

### 🔴 Critical Issues
- **[Issue Name]** at [file.ts#L42](file.ts#L42)
  - **Risk:** Description of what could go wrong
  - **Fix:** Recommended remediation

### 🟠 High Issues
- **[Issue Name]** at [file.ts#L15](file.ts#L15)
  - **Risk:** Description of vulnerability
  - **Fix:** Recommended remediation

### 🟡 Medium Issues
- **[Issue Name]** at [file.ts#L88](file.ts#L88)
  - **Risk:** Description of concern
  - **Fix:** Recommended remediation

### 🟢 Low Issues / Recommendations
- Consider [improvement suggestion]

### ✅ Passed Checks
- Mass assignment protection ✓
- Input validation ✓
- Authentication checks ✓

Secure Patterns for This Project

Express Route Security

// ✅ Correct: Whitelist fields explicitly
const { title, description, priorityId } = req.body
await prisma.task.create({
  data: { title, description, priorityId }
})

// ❌ Wrong: Mass assignment vulnerability
await prisma.task.create({ data: req.body })

Vue Template Security

<!-- ✅ Correct: Text interpolation (auto-escaped) -->
<p>{{ userInput }}</p>

<!-- ❌ Wrong: XSS vulnerability -->
<p v-html="userInput"></p>

Input Validation

// ✅ Correct: Validate before use
if (!title || typeof title !== 'string' || title.length > 200) {
  return res.status(400).json({ error: 'Invalid title' })
}

// ❌ Wrong: Trust user input
await prisma.task.create({ data: { title: req.body.title } })

AI-Generated Code Review

// ⚠️ AI suggested this package - VERIFY before installing:
// 1. Check npm: https://www.npmjs.com/package/fast-csv-parser
// 2. Verify downloads (>10k weekly), maintainer, last update
// 3. Check for known vulnerabilities: npm audit
import { parse } from 'fast-csv-parser'

// ⚠️ AI generated complex logic - REVIEW carefully:
// - Does this match requirements?
// - Are edge cases handled?
// - Is error handling complete?

MCP Configuration Security

// ✅ Correct: Use environment variable expansion
{
  "env": {
    "API_TOKEN": "${env:GITHUB_TOKEN}"
  }
}

// ❌ Wrong: Hardcoded credentials
{
  "env": {
    "API_TOKEN": "ghp_xxxxxxxxxxxx"
  }
}

Reference Documentation

For secure implementation patterns and comprehensive security guidelines, see:

• Backend Routes Instructions – Secure Express/Prisma patterns
• Security Guide – AI attack surface and operational checklist

MCP Risk Scoring Framework

Use this quantitative model to evaluate MCP server and tool risks:

$$R_{total} = \sum_{tool=1}^{n} (A_{tool} \times S_{tool} \times D_{tool})$$

Scoring Factors

Risk Thresholds

Capability-based overrides:

Even if a server's numeric score falls into the 0.0–0.25 "Auto-approve eligible" band, require at least per-session approval when:

• It can perform any non-read-only action (create/update/delete/execute/generate), or
• It can access internal, customer, or otherwise non-public data, or
• It is vendor-hosted and has broad access to your project or workspace.

These capability-based overrides ensure that powerful or data-sensitive integrations are never fully auto-approved, even with a low numeric risk score.

Example: Project MCP Servers